使用LoFTR模型进行图像配准、重叠区提取

embedded/2024/9/23 20:44:59/

LoFTR模型源自2021年CVPR提出的一篇论文LoFTR: Detector-Free Local Feature Matching with Transformers，其基于pytorch实现图像配准，与基于superpoint+superglue的方法不同，
是一个端到端的图像配准方法。与LoFTR官方库相关的有loftr2onnx库，整体来说loftr2onnx库使用更方便，效果更好。但loftr2onnx转出的onnx模型是有问题的，不能使用。

项目地址：https://github.com/zju3dv/LoFTR
项目地址2：https://github.com/oooooha/loftr2onnx
demo体验：https://huggingface.co/spaces/kornia/Kornia-LoFTR

1、web demo体验

访问 https://huggingface.co/spaces/kornia/Kornia-LoFTR

点配准的效果如下所示，可以看到与sp+sg方法相比，点对的平行关系直观上要好很多。
在这里插入图片描述

2、使用LoFTR项目

2.1 下载代码

打开https://github.com/zju3dv/LoFTR 下载代码
在这里插入图片描述

2.2 安装依赖项

安装项目使用：解压代码，然后在终端进入目录，执行pip install -r .\requirements.txt
在这里插入图片描述

2.3 下载模型

https://drive.google.com/drive/folders/1DOcOPZb3-5cWxLqn256AhwUVjBPifhuf 下载权重。如果需要训练，可以下载训练数据与从测试数据。
在这里插入图片描述
将权重的压缩文件解压后放到项目根目录下
在这里插入图片描述

2.4 初步使用

将 https://hpg123.blog.csdn.net/article/details/124824892 中章节3的代码（对应标题为 superpoint中read_img_as_tensor函数）保存为imgutils.py

from src.loftr import LoFTR, default_cfg
import torch
from imgutils import *
import time
# Initialize LoFTR
matcher = LoFTR(config=default_cfg)
matcher.load_state_dict(torch.load("weights/indoor_ds_new.ckpt")['state_dict'])
matcher = matcher.eval().cuda()p1=r'C:\Users\hpg\Pictures\t1.jpg'
p2=r'C:\Users\hpg\Pictures\t2.jpg'
t1,im1=read_img_as_tensor(p1,(384,384))
t2,im2=read_img_as_tensor(p2,(384,384))
batch = {'image0': t1, 'image1': t2}
# Inference
with torch.no_grad():matcher(batch)    # t0=time.time()times=10for i in range(times):matcher(batch)rt1=time.time()-t0rt1=rt1/timesmkpts0 = batch['mkpts0_f'].cpu().numpy()mkpts1 = batch['mkpts1_f'].cpu().numpy()mconf = batch['mconf'].cpu().numpy()print(f'运行时间：{rt1:.4f}',mkpts0.shape,mkpts1.shape,mconf)

代码运行效果如下所示，可以看到一个图片需要0.19s（笔记本，1060显卡），换成台式机3060显卡，预计在0.05s左右一张图

运行时间：0.1933 (32, 2) (32, 2) [0.22855578 0.21740437 0.34927088 0.28389925 0.27157754 0.269668280.22636016 0.22058277 0.20475665 0.20878278 0.22838292 0.254485850.27047077 0.34403533 0.22612476 0.2044811  0.26239234 0.327975540.2263804  0.26544347 0.3401669  0.39336586 0.3473139  0.282306940.23061718 0.23949552 0.46178365 0.3540019  0.5322925  0.272002370.26731068 0.39827508]

3、使用loftr2onnx库进行配准

3.1 基本使用

也可以基于 https://github.com/oooooha/loftr2onnx 项目进行图像配准
在这里插入图片描述
使用代码如下


from loftr_wrapper import LoFTRWrapper as LoFTR
import torch
from imgutils import *
import time
# Initialize LoFTR
matcher = LoFTR()
matcher.load_state_dict(torch.load("weights/indoor_ds_new.ckpt")['state_dict'])
matcher = matcher.eval().cuda()p1=r'C:\Users\hpg\Pictures\t1.jpg'
p2=r'C:\Users\hpg\Pictures\t2.jpg'
t1,im1=read_img_as_tensor(p1,(384,384))
t2,im2=read_img_as_tensor(p2,(384,384))
# Inference
with torch.no_grad():result=matcher(t1,t2)    # t0=time.time()times=10for i in range(times):result=matcher(t1,t2)rt1=time.time()-t0rt1=rt1/timesmkpts0 = result['keypoints0'].cpu().numpy()mkpts1 = result['keypoints1'].cpu().numpy()mconf = result['confidence'].cpu().numpy()print(f'运行时间：{rt1:.4f}',mkpts0.shape,mkpts1.shape,mconf)

代码输出如下所示，可以看到与LoFTR项目的输出有所差异

运行时间：0.1925 (212, 2) (212, 2) [0.4566688  0.53420454 0.5319168  0.5320238  0.46744433 0.40682140.5363396  0.45674214 0.60001785 0.6576139  0.53006035 0.595909240.5725811  0.5505655  0.44364485 0.40315574 0.4293331  0.50609730.6550978  0.52451503 0.553644   0.63088214 0.6906601  0.616680740.4543735  0.4138872  0.4332955  0.47855106 0.60533136 0.67351430.7912271  0.7220486  0.75414115 0.75669855 0.60389113 0.403050660.71130437 0.6583284  0.5403245  0.5433615  0.40149704 0.66738440.4093839  0.5410701  0.51509964 0.42121148 0.68238974 0.552473960.5116625  0.8369319  0.53321654 0.5323315  0.5779519  0.647059260.43591025 0.40134645 0.4599252  0.46620858 0.6388375  0.83547580.515318   0.6521981  0.54744494 0.64528877 0.7466613  0.63595170.58179545 0.4587202  0.4856584  0.42029297 0.43322447 0.432207580.6896481  0.79645556 0.5817581  0.75245494 0.5786756  0.72515590.814531   0.49031648 0.46484298 0.54241467 0.5943087  0.72451150.6457875  0.8097793  0.7199513  0.49220178 0.5443373  0.40861040.5046131  0.7193697  0.6752727  0.41796637 0.5513792  0.70874180.7779165  0.75016826 0.68525094 0.58962977 0.6315668  0.49130850.56355244 0.41288543 0.52281946 0.42782715 0.43921712 0.52160180.5566503  0.78442967 0.6013023  0.42023212 0.43102428 0.615640640.40717542 0.49634054 0.45509326 0.4511342  0.41775596 0.558971760.56803375 0.6018254  0.71239305 0.44001386 0.43651453 0.69477330.8648205  0.4988858  0.40208712 0.71607304 0.9030141  0.55438260.49472648 0.5359598  0.74733096 0.6617334  0.7066015  0.7256770.43446922 0.5126569  0.52367914 0.45096788 0.4248741  0.432852750.723374   0.86523044 0.65740126 0.427191   0.4776224  0.48018260.4530296  0.4275035  0.527438   0.52301216 0.58992577 0.417273430.48609605 0.7365703  0.6339512  0.6379226  0.4489899  0.413250480.5010124  0.49238032 0.57079905 0.62783945 0.5092921  0.57263870.60590863 0.44714844 0.6284152  0.40801758 0.40126294 0.42214190.52245826 0.70989937 0.49206337 0.553483   0.4956581  0.41806970.6228596  0.6543849  0.7747963  0.61180156 0.60290194 0.54211940.6149054  0.48783877 0.40048426 0.47044232 0.40145218 0.423598560.68902797 0.44713116 0.84827214 0.48961237 0.6137104  0.77524260.7184252  0.71058017 0.47483382 0.7151901  0.78853625 0.669882540.7502565  0.42592585 0.49173304 0.4657402  0.59592575 0.428502770.4645101  0.5070625 ]

3.2 差异分析

1、这主要是loftr2onnx项目通过loftr_wrapper对LoFTR的forward流程进行了调整。

#!/usr/bin/env python
import copy
import os
import sys
from typing import Any, Dictimport torch
from einops.einops import rearrange_CURRENT_DIR = os.path.dirname(os.path.realpath(__file__))
sys.path.append(os.path.join(_CURRENT_DIR, "LoFTR"))from loftr import LoFTR, default_cfgDEFAULT_CFG = copy.deepcopy(default_cfg)
DEFAULT_CFG["coarse"]["temp_bug_fix"] = Trueclass LoFTRWrapper(LoFTR):def __init__(self,config: Dict[str, Any] = DEFAULT_CFG,):LoFTR.__init__(self, config)def forward(self,image0: torch.Tensor,image1: torch.Tensor,) -> Dict[str, torch.Tensor]:data = {"image0": image0,"image1": image1,}del image0, image1data.update({"bs": data["image0"].size(0),"hw0_i": data["image0"].shape[2:],"hw1_i": data["image1"].shape[2:],})if data["hw0_i"] == data["hw1_i"]:  # faster & better BN convergencefeats_c, feats_f = self.backbone(torch.cat([data["image0"], data["image1"]], dim=0))(feat_c0, feat_c1), (feat_f0, feat_f1) = feats_c.split(data["bs"]), feats_f.split(data["bs"])else:  # handle different input shapes(feat_c0, feat_f0), (feat_c1, feat_f1) = self.backbone(data["image0"]), self.backbone(data["image1"])data.update({"hw0_c": feat_c0.shape[2:],"hw1_c": feat_c1.shape[2:],"hw0_f": feat_f0.shape[2:],"hw1_f": feat_f1.shape[2:],})# 2. coarse-level loftr module# add featmap with positional encoding, then flatten it to sequence [N, HW, C]feat_c0 = rearrange(self.pos_encoding(feat_c0), "n c h w -> n (h w) c")feat_c1 = rearrange(self.pos_encoding(feat_c1), "n c h w -> n (h w) c")mask_c0 = mask_c1 = None  # mask is useful in trainingif "mask0" in data:mask_c0, mask_c1 = data["mask0"].flatten(-2), data["mask1"].flatten(-2)feat_c0, feat_c1 = self.loftr_coarse(feat_c0, feat_c1, mask_c0, mask_c1)# 3. match coarse-levelself.coarse_matching(feat_c0, feat_c1, data, mask_c0=mask_c0, mask_c1=mask_c1)# 4. fine-level refinementfeat_f0_unfold, feat_f1_unfold = self.fine_preprocess(feat_f0, feat_f1, feat_c0, feat_c1, data)if feat_f0_unfold.size(0) != 0:  # at least one coarse level predictedfeat_f0_unfold, feat_f1_unfold = self.loftr_fine(feat_f0_unfold, feat_f1_unfold)# 5. match fine-levelself.fine_matching(feat_f0_unfold, feat_f1_unfold, data)rename_keys: Dict[str, str] = {"mkpts0_f": "keypoints0","mkpts1_f": "keypoints1","mconf": "confidence",}out: Dict[str, torch.Tensor] = {}for k, v in rename_keys.items():_d = data[k]if isinstance(_d, torch.Tensor):out[v] = _delse:raise TypeError(f"Expected torch.Tensor for item `{k}`. Gotcha {type(_d)}")del datareturn out

2、然后cfg或许有不同

loftr2onnx中的默认配置在loftr\utils\cvpr_ds_config.py中，如下所示。可以看到 _CN.MATCH_COARSE.THR与_CN.MATCH_COARSE.BORDER_RM 是做过修改的，与默认值不同

from yacs.config import CfgNode as CNdef lower_config(yacs_cfg):if not isinstance(yacs_cfg, CN):return yacs_cfgreturn {k.lower(): lower_config(v) for k, v in yacs_cfg.items()}_CN = CN()
_CN.BACKBONE_TYPE = 'ResNetFPN'
_CN.RESOLUTION = (8, 2)  # options: [(8, 2), (16, 4)]
_CN.FINE_WINDOW_SIZE = 5  # window_size in fine_level, must be odd
_CN.FINE_CONCAT_COARSE_FEAT = True# 1. LoFTR-backbone (local feature CNN) config
_CN.RESNETFPN = CN()
_CN.RESNETFPN.INITIAL_DIM = 128
_CN.RESNETFPN.BLOCK_DIMS = [128, 196, 256]  # s1, s2, s3# 2. LoFTR-coarse module config
_CN.COARSE = CN()
_CN.COARSE.D_MODEL = 256
_CN.COARSE.D_FFN = 256
_CN.COARSE.NHEAD = 8
_CN.COARSE.LAYER_NAMES = ['self', 'cross'] * 4
_CN.COARSE.ATTENTION = 'linear'  # options: ['linear', 'full']
_CN.COARSE.TEMP_BUG_FIX = False# 3. Coarse-Matching config
_CN.MATCH_COARSE = CN()
_CN.MATCH_COARSE.THR = 0.4  # thresh default=0.2
_CN.MATCH_COARSE.BORDER_RM = 4  # border default=2
_CN.MATCH_COARSE.MATCH_TYPE = 'dual_softmax'  # options: ['dual_softmax, 'sinkhorn']
_CN.MATCH_COARSE.DSMAX_TEMPERATURE = 0.1
_CN.MATCH_COARSE.SKH_ITERS = 3
_CN.MATCH_COARSE.SKH_INIT_BIN_SCORE = 1.0
_CN.MATCH_COARSE.SKH_PREFILTER = True
_CN.MATCH_COARSE.TRAIN_COARSE_PERCENT = 0.4  # training tricks: save GPU memory
_CN.MATCH_COARSE.TRAIN_PAD_NUM_GT_MIN = 200  # training tricks: avoid DDP deadlock# 4. LoFTR-fine module config
_CN.FINE = CN()
_CN.FINE.D_MODEL = 128
_CN.FINE.D_FFN = 128
_CN.FINE.NHEAD = 8
_CN.FINE.LAYER_NAMES = ['self', 'cross'] * 1
_CN.FINE.ATTENTION = 'linear'default_cfg = lower_config(_CN)

4、基于LoFTR提取重叠区域

4.1 使用LoFTR库

这里的imgutils 与前文说明的有细微区别，打开https://blog.csdn.net/a486259/article/details/124824892 将章节1的代码保存为imgutils.py即可

from src.loftr import LoFTR, default_cfg
import torch
from imgutils import *
import time
# Initialize LoFTR
matcher = LoFTR(config=default_cfg)
matcher.load_state_dict(torch.load("weights/indoor_ds_new.ckpt")['state_dict'])
matcher = matcher.eval().cuda()p1=r'C:\Users\hpg\Pictures\t1.jpg'
p2=r'C:\Users\hpg\Pictures\t2.jpg'
t1,im1=read_img_as_tensor_gray(p1,(384,384))
t2,im2=read_img_as_tensor_gray(p2,(384,384))
batch = {'image0': t1, 'image1': t2}
# Inference
with torch.no_grad():t0=time.time()times=1for i in range(times):matcher(batch)rt1=time.time()-t0rt1=rt1/timesmkpts0 = batch['mkpts0_f'].cpu().numpy()mkpts1 = batch['mkpts1_f'].cpu().numpy()confidence = batch['mconf'].cpu().numpy()print(f'运行时间：{rt1:.4f}',mkpts0.shape,mkpts1.shape)import cv2 as cv
pt_num = mkpts0.shape[0]
im_dst,im_res=im1,im2
img = np.zeros((max(im_dst.shape[0], im_res.shape[0]), im_dst.shape[1]+im_res.shape[1]+10,3))
img[:,:im_res.shape[0],]=im_dst
img[:,-im_res.shape[0]:]=im_res
img=img.astype(np.uint8)
match_threshold=0.6
for i in range(0, pt_num):if (confidence[i] > match_threshold):pt0 = mkpts0[i].to('cpu').numpy().astype(np.int32)pt1 = mkpts1[i].to('cpu').numpy().astype(np.int32)#cv.circle(img, (pt0[0], pt0[1]), 1, (0, 0, 255), 2)#cv.circle(img, (pt1[0], pt1[1]+650), (0, 0, 255), 2)cv.line(img, pt0, (pt1[0]+im_res.shape[0], pt1[1]), (0, 255, 0), 1)
myimshow( img,size=12)import cv2
def getGoodMatchPoint(mkpts0, mkpts1, confidence,  match_threshold:float=0.5):n = min(mkpts0.size(0), mkpts1.size(0))srcImage1_matchedKPs, srcImage2_matchedKPs=[],[]if (match_threshold > 1 or match_threshold < 0):print("match_threshold error!")for i in range(n):kp0 = mkpts0[i]kp1 = mkpts1[i]pt0=(kp0[0].item(),kp0[1].item());pt1=(kp1[0].item(),kp1[1].item());c = confidence[i].item();if (c > match_threshold):srcImage1_matchedKPs.append(pt0);srcImage2_matchedKPs.append(pt1);return np.array(srcImage1_matchedKPs),np.array(srcImage2_matchedKPs)
pts_src, pts_dst=getGoodMatchPoint(mkpts0, mkpts1, confidence)h1, status = cv2.findHomography(pts_src, pts_dst, cv.RANSAC, 8)
im_out1 = cv2.warpPerspective(im_dst, h1, (im_dst.shape[1],im_dst.shape[0]))
im_out2 = cv2.warpPerspective(im_res, h1, (im_dst.shape[1],im_dst.shape[0]),16)
#这里 im_res和im_out1是严格配准的状态
myimshowsCL([im_dst,im_out1,im_res,im_out2],rows=2,cols=2, size=6)

在这里插入图片描述
代码运行报错，因为匹配到的点太少了，无法计算转化矩阵提取重叠区

4.2 使用loftr2onnx库

这里的imgutils 与前文说明的有细微区别，打开https://blog.csdn.net/a486259/article/details/124824892 将章节1的代码保存为imgutils.py即可

使用代码如下


from loftr_wrapper import LoFTRWrapper as LoFTR
import torch
from imgutils import *
import time
# Initialize LoFTR
matcher = LoFTR()
matcher.load_state_dict(torch.load("weights/indoor_ds_new.ckpt")['state_dict'])
matcher = matcher.eval().cuda()p1=r'C:\Users\hpg\Pictures\t1.jpg'
p2=r'C:\Users\hpg\Pictures\t2.jpg'
t1,im1=read_img_as_tensor_gray(p1,(384,384))
t2,im2=read_img_as_tensor_gray(p2,(384,384))
# Inference
with torch.no_grad():#result=matcher(t1,t2)    # t0=time.time()times=1for i in range(times):result=matcher(t1,t2)rt1=time.time()-t0rt1=rt1/timesmkpts0 = result['keypoints0']#.cpu().numpy()mkpts1 = result['keypoints1']#.cpu().numpy()confidence = result['confidence']#.cpu().numpy()print(f'运行时间：{rt1:.4f}',mkpts0.shape,mkpts1.shape,confidence)import cv2 as cv
pt_num = mkpts0.shape[0]
im_dst,im_res=im1,im2
img = np.zeros((max(im_dst.shape[0], im_res.shape[0]), im_dst.shape[1]+im_res.shape[1]+10,3))
img[:,:im_res.shape[0],]=im_dst
img[:,-im_res.shape[0]:]=im_res
img=img.astype(np.uint8)
match_threshold=0.01
for i in range(0, pt_num):if (confidence[i] > match_threshold):pt0 = mkpts0[i].to('cpu').numpy().astype(np.int32)pt1 = mkpts1[i].to('cpu').numpy().astype(np.int32)#cv.circle(img, (pt0[0], pt0[1]), 1, (0, 0, 255), 2)#cv.circle(img, (pt1[0], pt1[1]+650), (0, 0, 255), 2)cv.line(img, tuple(pt0.tolist()), (pt1[0]+im_res.shape[0], pt1[1]), (0, 255, 0), 1)
myimshow( img,size=12)import cv2
def getGoodMatchPoint(mkpts0, mkpts1, confidence,  match_threshold:float=0.5):n = min(mkpts0.size(0), mkpts1.size(0))srcImage1_matchedKPs, srcImage2_matchedKPs=[],[]if (match_threshold > 1 or match_threshold < 0):print("match_threshold error!")for i in range(n):kp0 = mkpts0[i]kp1 = mkpts1[i]pt0=(kp0[0].item(),kp0[1].item());pt1=(kp1[0].item(),kp1[1].item());c = confidence[i].item();if (c > match_threshold):srcImage1_matchedKPs.append(pt0);srcImage2_matchedKPs.append(pt1);return np.array(srcImage1_matchedKPs),np.array(srcImage2_matchedKPs)
pts_src, pts_dst=getGoodMatchPoint(mkpts0, mkpts1, confidence)h1, status = cv2.findHomography(pts_src, pts_dst, cv.RANSAC, 4)
# im_dst=im_dst.astype(np.float32)/255
# im_res=im_res.astype(np.float32)/255
print(im_dst.shape,im_dst.dtype,im_dst.max(),im_res.shape,im_res.dtype,im_res.max(),h1)
im_out1 = cv2.warpPerspective(im_dst, h1, (im_dst.shape[1],im_dst.shape[0]))
im_out2 = cv2.warpPerspective(im_res, h1, (im_dst.shape[1],im_dst.shape[0]),16)
#这里 im_res和im_out1是严格配准的状态
myimshowsCL([im_dst,im_out1,im_res,im_out2],rows=2,cols=2, size=6)

提取的点对关系如下所示
在这里插入图片描述
提取出的重叠区域
在这里插入图片描述


http://www.ppmy.cn/embedded/56718.html

相关文章

gdb-dashboard：用Python重塑GDB调试体验

gdb-dashboard：用Python重塑GDB调试体验

gdb-dashboard&#xff1b;一目了然的GDB调试&#xff0c;尽在掌控之中- 精选真开源&#xff0c;释放新价值。 概览 gdb-dashboard是一个用Python编写的模块化视觉界面&#xff0c;为GNU Debugger&#xff08;GDB&#xff09;提供了一个现代化的工作空间。它通过集成多个面板和…
阅读更多...
人工智能写作对话系统源码 自然语言的处理能力 前后端分离 带完整的安装代码包以及搭建教程

人工智能写作对话系统源码 自然语言的处理能力 前后端分离 带完整的安装代码包以及搭建教程

系统概述 随着互联网信息爆炸式增长&#xff0c;用户对于高质量、个性化内容的需求日益增长&#xff0c;而传统的内容生成方式已难以满足这一需求。另一方面&#xff0c;深度学习和自然语言处理技术的突破性进展&#xff0c;为人机交互提供了新的可能。本项目正是在此背景下应…
阅读更多...
Linux 防火墙配置指南：firewalld 端口管理应用案例(二十个实列)

Linux 防火墙配置指南：firewalld 端口管理应用案例(二十个实列)

&#x1f3e1;作者主页&#xff1a;点击&#xff01; &#x1f427;Linux基础知识(初学)&#xff1a;点击&#xff01; &#x1f427;&#x1f427;Linux高级管理专栏&#xff1a;点击&#xff01; &#x1f510;Linux中firewalld防火墙&#xff1a;点击&#xff01; ⏰️…
阅读更多...
【BP时序预测】基于布谷鸟优化算法CS实现负荷数据预测单输入单输出附matlab代码

【BP时序预测】基于布谷鸟优化算法CS实现负荷数据预测单输入单输出附matlab代码

% 负荷数据预测单输入单输出&#xff08;BP时序预测&#xff09; % 使用布谷鸟优化算法实现 % 假设你已经有了输入数据和对应的输出数据 % 输入数据应该是一个矩阵&#xff0c;每一行代表一个样本&#xff0c;每一列代表一个特征 % 输出数据应该是一个列向量&#xff0c;每个元…
阅读更多...
el-table实现固定列，及解决固定列导致部分滚动条无法拖动的问题

el-table实现固定列，及解决固定列导致部分滚动条无法拖动的问题

一、el-table实现固定列 当数据量动态变化时&#xff0c;可以为 Table 设置一个最大高度。 通过设置max-height属性为 Table 指定最大高度。此时若表格所需的高度大于最大高度&#xff0c;则会显示一个滚动条。 <div class"zn-filter-table"><!-- 表格--…
阅读更多...
Redis基础教程（九）：redis有序集合

Redis基础教程（九）：redis有序集合

&#x1f49d;&#x1f49d;&#x1f49d;首先&#xff0c;欢迎各位来到我的博客&#xff0c;很高兴能够在这里和您见面&#xff01;希望您在这里不仅可以有所收获&#xff0c;同时也能感受到一份轻松欢乐的氛围&#xff0c;祝你生活愉快&#xff01; &#x1f49d;&#x1f49…
阅读更多...
拦截HTTP的多种方式

拦截HTTP的多种方式

部分场景下需要修改请求报文信息&#xff0c;可以利用 AOP 思维&#xff08;切面编程&#xff09;&#xff0c;对请求进行拦截处理。Web 中有见的几种发送请求的方式&#xff1a; XMLHttpRequestfetchwindow.navigator.sendBeaconnew Imageservice worker 针对这几种不同的场…
阅读更多...
数据结构-第八章（1.基本概念）

数据结构-第八章（1.基本概念）

注&#xff1a;这一章我花了两天半的时间过完&#xff0c;有点赶&#xff0c;对于一轮复习&#xff0c;我没有做8.6各部分内部排序算法的比较及应用。有点算留了坑&#xff0c;这也是我二轮在这一章的重点补充。 考纲内容 知识框架 复习提示 堆排序、快速排序和归并排序是这章…
阅读更多...
最新文章

Copyright @ 2022~2023